Add new export LLM config #11028
Add new export LLM config #11028
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11028
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 4372d61 with merge base c2aa614 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
|
@jackzhxng has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75263991 |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
Pull Request resolved: #11028 @imported-using-ghimport Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991/) ghstack-source-id: 288636930
|
This pull request was exported from Phabricator. Differential Revision: D75263991 |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75263991 |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75263991 |
Differential Revision: [D75263991](https://our.internmc.facebook.com/intern/diff/D75263991) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75263991 |
e46a59a
into
gh/jackzhxng/10/base
This PR was created by the merge bot to help merge the original PR into the main branch. ghstack PR number: #11028 by @jackzhxng ^ Please use this as the source of truth for the PR details, comments, and reviews ghstack PR base: https://github.com/pytorch/executorch/tree/gh/jackzhxng/10/base ghstack PR head: https://github.com/pytorch/executorch/tree/gh/jackzhxng/10/head Merge bot PR base: https://github.com/pytorch/executorch/tree/main Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/jackzhxng/10/orig @diff-train-skip-merge Co-authored-by: Jack Zhang <[email protected]>
| SMOLLM2 = "smollm2" | ||
|
|
||
|
|
||
| class PreqMode(str, Enum): |
There was a problem hiding this comment.
worth tagging as deprecated
| model_class: Which model to to export. | ||
| params: Model parameters, such as n_layers, hidden_size, etc. | ||
| If left empty will use defaults specified in model_args.py. | ||
| checkpoint: Path to the checkpoint file. |
There was a problem hiding this comment.
can this be hf path as well?
There was a problem hiding this comment.
No, but at the moment if you specify a non-llama model_class and don't specify checkpoint, it will download from HF. Worth adding this comment
| tokenizer_path: Path to the tokenizer file. | ||
| metadata: Json string containing metadata information. | ||
| e.g. '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}' | ||
| use_lora: Rank of the LoRA, if set to 0 then this means no LoRA. For use with QAT. |
There was a problem hiding this comment.
lora in our case is really tied to QAT model that we released, right? It is not independently applicable to any model? If so I think we want to tied this to QAT checkpoints specifically for llama3_2
There was a problem hiding this comment.
Is that the case? If so I'll add something to the post_init to verify and update the comment
| metadata: Json string containing metadata information. | ||
| e.g. '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}' | ||
| use_lora: Rank of the LoRA, if set to 0 then this means no LoRA. For use with QAT. | ||
| fairseq2: For legacy internal use cases, this is safe to ignore. |
There was a problem hiding this comment.
prefix with _ to differentiate between this and other supported params
| preq_mode: Legacy option to specify how prequantized weights are loaded. | ||
| Going forward, ExecuTorch supports loading weights prequantized through | ||
| TorchAo as-is, without any special handling. | ||
| preq_group_size: Legacy option to specify the group size of prequantized weights. |
There was a problem hiding this comment.
I thik we probably want to couple these together. If you are loading pre-quantized checkpoint that group size cannot be set independently, right? So maybe having a separate dataclass that captures all the params and maps it by name is better. ALthough I presume you have to keep this for BC?
There was a problem hiding this comment.
Yeah, just keeping for BC. I think ideally we are moving away from preq to tho right?
| @dataclass | ||
| class ModelConfig: | ||
| """ | ||
| Configurations not necessarily specific to the model, but are needed to |
There was a problem hiding this comment.
Then why call it modelconfig?
There was a problem hiding this comment.
Maybe call LoweringConfig? or ExportConfig although not all options are export related
There was a problem hiding this comment.
This is more like: what other modifications do I want to make to the model in eager that aren't specific to the model itself (e.g. NOT checkpoint, model architecture, tokenizer) and can be shared across different models.
| doesn't actually have anything to do with the kv_cache at the moment. | ||
| expand_rope_table: Temporary workaround to expand sin/cos table in head | ||
| dim to take vectorized path in optimized kernels. | ||
| use_attention_sink: Whether to use attention sink to support multi-round |
There was a problem hiding this comment.
I would also considering pruning options that we dont intend to support. attention sink at the moment does not have performant implementation and I would rather hide it somewhere than to expose it. Reduce maintenance burden
| max_context_length: Maximum of context for the model to remember. | ||
| output_dir: Output dir to save the exported .pte file to. | ||
| output_name: File name to override the exported .pte file. | ||
| so_library: Shared library to specify custom quantized operators. |
There was a problem hiding this comment.
I dont think we need so_library anymore. Remove it in follow up
| output_dir: Output dir to save the exported .pte file to. | ||
| output_name: File name to override the exported .pte file. | ||
| so_library: Shared library to specify custom quantized operators. | ||
| export_only: Whether to stop right after torch.export() and |
| XNNPACK_DYNAMIC = "xnnpack_dynamic" | ||
| XNNPACK_DYNAMIC_QC4 = "xnnpack_dynamic_qc4" |
There was a problem hiding this comment.
we probably can remove these two
| pt2e_quantize: Quantization mode using pt2e, which is an alternative | ||
| to TorchAo that uses backend-aware graph mode quantization rather | ||
| than source transformation quantization. |
There was a problem hiding this comment.
I would want to hide these details from users
| class CoreMLQuantize(str, Enum): | ||
| B4W = "b4w" | ||
| C4W = "c4w" |
There was a problem hiding this comment.
maybe favor this pattern over the ones in pt2e_quantizer
| class TestValidConstruction(unittest.TestCase): | ||
|
|
||
| def test_valid_llm_config(self): | ||
| LlmConfig( |
There was a problem hiding this comment.
are these configs constructible from json? I think that would e the best
Stack from ghstack (oldest at bottom):
Differential Revision: D75263991